AI Security 10
The collective cyberdefense letter's three asks: stop assuming status quo security is enough, put cyber-capable AI in defenders' hands, and fund it globally — including governments paying for essential services to get defensive AI
(letter · Andrew Curran's summary) `[RELEVANT]`
Explainer →
Sam Altman: "there is not much time to act… only an urgent and intense collective response will work"
(sama)
Explainer →
A quiet HN post asks security people to look at the Microduck GitHub: camera and mic reachable with no auth on the same network, and Wi-Fi credentials exposed
(HN 1 pt — low traffic, unverified · repo) `[RELEVANT]`
Explainer →
Five Primitives for Governing Autonomous AI Agents at Runtime — enterprise identity was built for humans and long-lived services, and agents break it in three specific ways
(arxiv) `[RELEVANT]`
Explainer →
Risks and Controls for Multi-Agent Systems: a framework for what happens when agents from different organisations interact
(arxiv) `[RELEVANT]`
Explainer →
INTENT-AS-A-TOOL: make the agent declare its intent as a tool call, and agentic misalignment becomes trackable
(arxiv) `[RELEVANT]`
Explainer →
AnTrap: Android GUI agents are not robust against runtime anomalies — a dynamic adversarial environment for testing them
(arxiv) `[RELEVANT]`
Explainer →
The Compaction Cliff: when an agent's context overflows, the safety rule and the episodic log get summarised at the same rate — but only the rule needs exact wording to stay enforceable
(arxiv) `[RELEVANT]`
Explainer →
Persona-Execution Separation: let the persona evolve freely, keep execution audited — an architecture pattern for governed LLM agents
(arxiv)
Explainer →
Mollick on the METR report: it is good and important, but people are ascribing far too many human motivations to the agents involved
(emollick)
Explainer →
AI Agents 12
Anthropic's Model Hardware Standard: reduce weeks of lab-instrument integration to hours, and let agents run round-the-clock experiments that recover from hardware errors on their own
(Anthropic)
Explainer →
Terminal-Bench-Science: 70 workflows drawn from researchers' own work, and the best model — Claude Opus 5 — resolves 30%
(announcement · HN 34 pts)
Explainer →
JIT-Agent: the harness, not the model, often dominates agent capability — so evolve the harness just-in-time (47↑)
(arxiv)
Explainer →
The Handoff Tax: every time you escalate to a stronger model or downshift to a cheaper one mid-run, the receiver pays a measurable cost to continue someone else's trajectory (12↑)
(arxiv)
Explainer →
SWE Refactor Bench: can a coding agent complete a long-horizon, whole-repository stack migration? (13↑)
(arxiv)
Explainer →
WikiSkill: compile agent experience into persistent knowledge instead of rediscovering skills every run
(arxiv)
Explainer →
Autonomous mathematical discovery in "the Station" — an open world where agents from different model families pursue a shared research goal with no coordinator
(arxiv · _akhaliq)
Explainer →
SimVerity: how much evidence does a simulated pass actually give you about physical deployment?
(arxiv)
Explainer →
ReproAgent: contract-guided paper-to-code reproduction, where the specification is split between the paper and the artifacts
(arxiv)
Explainer →
SPECMINE: a large-scale corpus of spec-driven development artifacts — the practice where a natural-language spec, usually AI-drafted and human-curated, drives the agent
(arxiv)
Explainer →
New research on agentic shopping: you cannot reliably predict or influence what an agent buys — page viewing order and memories change its preferences unpredictably
(emollick)
Explainer →
"Harness Engineering" hits HN, and the thread mostly argues about whether harness / skills / agents are real concepts or just new words
(post · HN 118 pts, 42 comments)
Explainer →
HN Threads 12
Cloudflare cut 1.1.1.1's DNS cache footprint by more than half — 100TB of RAM freed, insert throughput up 43%, lookup latency down 19%
(Cloudflare · HN 557 pts, 159 comments)
Explainer →
"Six months of writing code exclusively with agents" — and the top comment is about typing being the bottleneck, not thinking
(post · HN 66 pts, 98 comments)
Explainer →
"Software engineering is about managing complexity" — the argument is old, the fight in the comments is new
(post · HN 112 pts, 79 comments)
Explainer →
Silicon Valley is in denial about the backlash — 116 comments that turn into a conversation about purpose
(Blood in the Machine · HN 93 pts, 116 comments)
Explainer →
"The Teaser Period: why the AI boom is hitting a reset wall" — take-or-pay compute obligations as the senior tranche above OpenAI's equity
(post · HN 86 pts, 78 comments)
Explainer →
Amazon now hides customer reviews from signed-out visitors — anti-scraping, or keeping the data in?
(HN 69 pts, 31 comments)
Explainer →
507 Mechanical Movements, still one of the best things on the internet
(site · HN 485 pts, 67 comments)
Explainer →
Decompiling a Nintendo 64 game in 84 days
(post · HN 189 pts, 121 comments)
Explainer →
The story of Suica, Japan's first IC transit card
(TokyoDev · HN 196 pts, 168 comments)
Explainer →
Yayoi Kusama has died
(BBC · HN 108 pts)
Explainer →
Peter Cullen, the voice of Optimus Prime, has died
(Hollywood Reporter · HN 103 pts)
Explainer →
The US government designates the host of noblogs.org a "global terrorist"
(CrimethInc · HN 214 pts, 120 comments)
Explainer →